JMIR Publications

ABSTRACT

Background:

Posttraumatic epilepsy (PTE) is a common sequela after traumatic brain injury (TBI) and increases the morbidity and mortality. Identifying patients at high risk of PTE is necessary for their better treatment. Risk prediction model of artificial neural network (ANN) had been reported and outperformed than conventional model, while the ANN prediction model for PTE is lacking.

Objective:

We aimed to train and validate ANN model to predict risk of PTE.

Methods:

1301 individuals diagnosed as TBI between January 1, 2011 and December 31, 2017 were enrolled from West China Hospital. Demographic data, clinical manifestations and radiological results of these patients were collected. We used a 5-fold cross-validation approach to train and test the ANN model to avoid overfitting. 21 independent variables employed as the input neurons in the ANN models, using back-propagation algorithm to minimize the loss function. We obtained sensitivity, specificity, accuracy of each ANN model from the 5 rounds of cross-validation and compared the accuracy with a nomogram prediction model built in our former work (data unpublished) based on the same population. The performance of this model was evaluated in patients registered at Chengdu Shang Jin Nan Fu Hospital (testing 1 cohort, n = 413) and Sichuan Provincial People’s Hospital (testing 2 cohort, n = 421) between January 1, 2013, and March 1, 2015.

Results:

Ultimately, we selected the best back-propagation 3-layer ANN prediction model, which included 21, 43, and 1 neuron in the input, hidden, and output layers, respectively. The area under (AUC) the receiver of receiver operating characteristics (ROC) curve in the training cohort was 0.907 (95% CI: 0.889-0.924)), testing 1 cohort was 0.867 (95% CI: 0.842-0.893) and testing 2 cohort was 0.859 (95% CI: 0.826-0.890), results showed the ANN model performed favorable. The average precision (AP) was 0.557 (95% CI: 0.510-0.620) in training cohort, 0.470 (95% CI: 0.414-0.526) in testing 1 cohort and 0.344 (95% CI: 0.287-0.401) in testing 2 cohort. The sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) in the training cohort (testing 1 cohort and testing 2 cohort) were 0.80 (0.83 and 0.80), 0.86 (0.80 and 0.84), 91% (85% and 78%) and 86% (80% and 83%). When calibrated this ANN model, the Brier score was 0.121 in the testing 1 cohort and 0.127 in the testing 2 cohort. Compared with nomogram model in our former work, ANN prediction model had a higher accuracy (P = .04).

Conclusions:

This study suggests that the ANN model could predict the risk of PTE in individuals after TBI and outperformed than that built by conventional statistical methods. While the calibration of this model was a bit poor, we need to calibrate it on large-sample-size set and further refine this model.